LV-ROVER-MLT: Low-Resource Maltese OCR by Synthetic Fine-Tuning and Multi-Stream Arbitration

2026-07-25 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Natural Language Processing · Depth: Expert, long

Summary

LV-ROVER-MLT is a novel system designed for low-resource Maltese Optical Character Recognition (OCR), addressing the scarcity of large-scale real labelled PDF corpora for the language. The system employs a synthetic training pipeline and a 5-stream Tesseract LV-ROVER ensemble, achieving a character error rate (CER) of 0.00700 on a 422-paragraph Maltese benchmark. This represents a 70 percent reduction compared to a fine-tuned-Tesseract baseline of CER 0.0234. The ensemble recognition alone improved CER by 44 percent to 0.01317, with a five-stage post-processing chain further reducing errors. Key components include a reproducible Maltese paragraph synthesis pipeline utilizing the 467M-token korpus_malti corpus and 68 validated fonts, along with a dual-CER reporting protocol to distinguish recognition gains from typographic alignment. The system operates on CPU, without GPU, emphasizing an ensemble-of-small-models approach.

Key takeaway

For machine learning engineers developing OCR for low-resource languages, prioritize synthetic data generation pipelines. Validate fonts and simulate realistic document degradations. Implement multi-stream ensemble recognition, varying language chains and image scales, to significantly reduce character error rates. Crucially, distinguish true recognition improvements from label-convention alignment. Incorporate diacritic-preserving logic to maintain linguistic accuracy.

Key insights

Synthetic data and multi-stream ensembles significantly improve low-resource OCR, especially for diacritic-rich languages.

Principles

• Synthetic data pipelines enable low-resource OCR.
• Diverse ensembles improve recognition by combining varied errors.
• Separate recognition gains from typographic normalization.

Method

The LV-ROVER system uses a 5-stream Tesseract LSTM ensemble, voted per word with a soft Maltese lexicon, followed by a five-stage label-convention normalisation chain and a rule-based line joiner.

In practice

• Synthesize data with font validation and realistic augmentations.
• Use diverse Tesseract streams (language, scale, fine-tuned).
• Apply diacritic-preserving edit-distance bounds in voting.

Topics

• Low-Resource OCR
• Maltese Language Processing
• Synthetic Data Generation
• Tesseract LSTM
• Ensemble OCR
• Diacritic Handling

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• KazakhOCR: A Synthetic Benchmark for Evaluating Multimodal Models in Low-Resource Kazakh Script OCR — Current MLLMs struggle with OCR and language identification for low-resource Kazakh scripts, especially Arabic, performing worse than traditional OCR.
• BLPR: Robust License Plate Recognition under Viewpoint and Illumination Variations via Confidence-Driven VLM Fallback — A modular LPDR system uses synthetic data, adaptive preprocessing, and a VLM fallback for robust recognition in challenging conditions.
• Advancing WordArt-Oriented Scene Text Recognition: Datasets and Methods — Recognizing highly customized WordArt demands specialized datasets and models beyond general Scene Text Recognition.
• Persian Pixel: A large-scale synthetic OCR dataset for Persian language — Synthetic data generation, exemplified by Persian Pixel, offers a scalable solution for OCR in typographically complex, low-resource languages.
• Why Nepali OCR Is Brutally Hard — And How We Finally Solved It — A hybrid deep learning pipeline overcomes complex script challenges to automate document data extraction and form pre-filling.
• VDE Bench: Evaluating The Capability of Image Editing Models to Modify Visual Documents — Image editing models require specialized benchmarks to accurately assess their performance on multilingual and densely textual visual documents.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.